The SimpleScalar Tool Set, Version 2.

0
Doug Burger

Todd M. Austin

Computer Sciences Department

University of Wisconsin-Madison
1210 West Dayton Street
Madison, Wisconsin 53706 USA

SimpleScalar LLC
2395 Timbercrest Court
Ann Arbor, MI 48105

*Contact: info@simplescalar.com
http://www.simplescalar.com
easy annotation of instructions, without requiring a retargeted
compiler for incremental changes. The instruction definition
method, along with the ported GNU tools, makes new simulators
easy to write, and the old ones even simpler to extend. Finally,
the simulators have been aggressively tuned for performance,
and can run codes approaching real sizes in tractable amounts
of time. On a 200-MHz Pentium Pro, the fastest, least detailed
simulator simulates about four million machine cycles per second, whereas the most detailed processor simulator simulates
about 150,000 per second.
The current release (version 2.0) of the tools is a major
improvement over the previous release. Compared to version 1.0
[2], this release includes better documentation, enhanced performance, compatibility with more platforms, precompiled SPEC95
SimpleScalar binaries, cleaner interfaces, two new processor
simulators, option and statistic management packages, a sourcelevel debugger (DLite!) and a tool to trace the out-of-order pipeline.
The rest of this document contains information about obtaining, installing, running, using, and modifying the tool set. In
Section 2 we provide a detailed procedure for downloading the
release, installing it, and getting it up and running. In Section 3,
we describe the SimpleScalar architecture and details about the
target (simulated) system. In Section 4, we describe the SimpleScalar processor simulators and discuss their internal workings. In
Section 5, we describe two tools that enhance the utility of the
tool set: a pipeline tracer and a source-level debugger (for stepping through the program being simulated). In Section 6, we provide the history of the tools development, describe current and
planned efforts to extend the tool set, and conclude. In
Appendix A and Appendix B contain detailed definitions of the
SimpleScalar instructions and system calls, respectively.

This report describes release 2.0 of the SimpleScalar tool set,

a suite of powerful computer simulation tools that provide both
detailed and high-performance simulation of modern microprocessors. The new release offers more tools and capabilities, precompiled binaries, cleaner interfaces, better documentation,
easier installation, improved portability, and higher performance. This report contains a complete description of the tool
set, including retrieval and installation instructions, a description of how to use the tools, a description of the target SimpleScalar architecture, and many details about the internals of the
tools and how to customize them. With this guide, the tool set can
be brought up and generating results in under an hour (on supported platforms).

1 Overview
Modern processors are incredibly complex marvels of engineering that are becoming increasingly hard to evaluate. This
report describes the SimpleScalar tool set (release 2.0), which
performs fast, flexible, and accurate simulation of modern processors that implement the SimpleScalar architecture (a close
derivative of the MIPS architecture [4]). The tool set takes binaries compiled for the SimpleScalar architecture and simulates
their execution on one of several provided processor simulators.
We provide sets of precompiled binaries (including SPEC95),
plus a modified version of GNU GCC (with associated utilities)
that allows you to compile your own SimpleScalar test binaries
from FORTRAN or C code.
The advantages of the SimpleScalar tools are high flexibility,
portability, extensibility, and performance. We include five execution-driven processor simulators in the release. They range
from an extremely fast functional simulator to a detailed, out-oforder issue, superscalar processor simulator that supports nonblocking caches and speculative execution.
The tool set is portable, requiring only that the GNU tools
may be installed on the host system. The tool set has been tested
extensively on many platforms (listed in Section 2). The tool set
is easily extensible. We designed the instruction set to support

2 Installation and Use

Authorized users may use and share SimpleScalar provided
that (1) the copyright notice must accompany all re-releases of
the tool set, and (2) third parties (i.e., you) are forbidden to place
any additional distribution restrictions on extensions to the tool
set that you release. The copyright notice can be found in the distribution directory as well as at the head of all simulator source
files. We have included the copyright here as well:
Copyright (C) 1994, 1995, 1996, 1997 by Todd M. Austin

These utilities are not required to run the simulators themselves, but is required to compile your own SimpleScalar
benchmark binaries (e.g. test programs other than the ones
we provide). The compressed file is 3 MB, the uncompressed file is 14 MB, and the build requires 52 MB.

simpletools.tar.gz - contains the retargeted GNU compiler

and library sources needed to build SimpleScalar benchmark binaries (GCC 2.6.3, glibc 1.0.9, and f2c), as well as
pre-built big- and little-endian versions of libc. This file is
needed only to build benchmarks, not to compile or run the
simulators. The tools are 11 MB compressed, 47 MB
uncompressed, and the full installation requires 70 MB.

simplebench.big.tar.gz - contains a set of the SPEC95

benchmark binaries, compiled to the SimpleScalar architecture running on a big-endian host. The binaries take under 5
MB compressed, and are 29 MB when uncompressed.

simplebench.little.tar.gz - same as above, except that the

binaries were compiled to the SimpleScalar architecture
running on a little-endian host.
Once you have selected the appropriate files, place the downloaded files into the desired target directory. If you obtained the
files with the .gz suffix, run the GNU decompress utility (gunzip). The files should now have a .tar suffix. To remove the
directories from the archive:

This tool set is distributed as is in the hope that it will be

useful. The tool set comes with no warranty, and no author or
distributor accepts any responsibility for the consequences of its
use.
Everyone is granted permission to copy, modify and redistribute this tool set under the following conditions:

This tool set is distributed for non-commercial use only.

Please contact the maintainer for restrictions applying to
commercial use of these tools.

Permission is granted to anyone to make or distribute copies of this tool set, either as received or modified, in any
medium, provided that all copyright notices, permission and
nonwarranty notices are preserved, and that the distributor
grants the recipient permission for further redistribution as
permitted by this document.

Permission is granted to distribute these tools in compiled

or executable form under the same conditions that apply for
source code, provided that either: (1) it is accompanied by
the corresponding machine-readable source code, or (2) it
is accompanied by a written offer, with no time limit, to give
anyone a machine-readable copy of the corresponding
source code in return for reimbursement of the cost of distribution. This written offer must permit verbatim duplication
by anyone, or (3) it is distributed by someone who received
only the executable form, and is accompanied by a copy of
the written offer of source code that they received concurrently.
The latest version of the SimpleScalar license can be found
at the following web address:

tar xf filename.tar

If you download and unpack all files, release, you should have
the following subdirectories with following contents:

simplesim-2.0 - the sources of the SimpleScalar processor

simulators, supporting scripts, and small test benchmarks. It
also holds precompiled binaries of the test benchmarks.

binutils-2.5.2 - the GNU binary utilities code, ported to the

SimpleScalar architecture.

ssbig-na-sstrix - the root directory for the tree in which the

big-endian SimpleScalar binary utilities and compiler tools
will be installed. The unpacked directories contain header
files and a pre-compiled copy of libc and a necessary object
file.

sslittle-na-sstrix - same as above, except that this directory

holds the little-endian versions of the SimpleScalar utilities.

gcc-2.6.3 - the GNU C compiler code, targeted toward the

SimpleScalar architecture.

glibc-1.09 - the GNU libraries code, ported to the SimpleScalar architecture.

f2c-1994.09.27 - the 1994 release of AT&T Bell Labs

FORTRAN to C translator code.

spec95-big - precompiled SimpleScalar SPEC95 benchmark binaries (big-endian version).

http://www.simplescalar.com/license.html

2.1 Obtaining the tools

The tools can either be obtained through the World Wide
Web, or by conventional ftp. For example, to get the file simplesim.tar.gz via the WWW, enter the URL:
www.simplescalar.com

and to obtain the same file with traditional ftp:

ftp ftp.simplescalar.com
user: anonymous
password: enter your e-mail address here
cd pub/simplescalar
get simplesim.tar

Note the tar.gz suffix: by requesting the file without the .gz
suffix, the ftp server uncompresses it automatically. To get the
compressed version, simply request the file with the .gz suffix.
The five distribution files in the directory (which are symbolic
links to the files containing the latest version of the tools) are:

simplesim.tar.gz - contains the simulator sources, the

instruction set definition macros, and test program source
and binaries. The directory is 1 MB compressed and 4 MB
uncompressed. When the simulators are built, the directory
(including object files) will require 11 MB. This file is
required for installation of the tool set.

simpleutils.tar.gz - contains the GNU binutils source (version 2.5.2), retargeted to the SimpleScalar architecture.

spec95-little - precompiled SimpleScalar SPEC95 benchmark binaries (little-endian version)

2.2 Installing and running Simplescalar

We depict a graphical overview of the tool set in Figure 1.
Benchmarks written in FORTRAN are converted to C using Bell
Labs f2c converter. Both benchmarks written in C and those
converted from FORTRAN are compiled using the SimpleScalar

FORTRAN
benchmark source

C
benchmark source

Simulator source
(e.g., sim-outorder.c)

f2c

SimpleScalar
GCC

Host C compiler

SimpleScalar
assembly
Simulator

SimpleScalar
GAS

SS libc.a
SS libm.a

RESULTS

Object files
Simplescalar
GLD

SS libF77.a

SimpleScalar
executables

Precompiled SS
binaries (test, SPEC95)

Figure 1. SimpleScalar tool set overview

version of GCC, which generates SimpleScalar assembly. The
SimpleScalar assembler and loader, along with the necessary
ported libraries, produce SimpleScalar executables that can then
be fed directly into one of the provided simulators. (The simulators themselves are compiled with the host platforms native
compiler; any ANSI C compiler will do).
If you use the precompiled SPEC95 binaries or the precompiled test programs, all you have to install is the simulator source
itself. If you wish to compile your own benchmarks, you will
have to install and build the GCC tree and optionally (recommended) the GNU binutils. If you wish to modify the support
libraries, you will have to install, modify, and build the glibc
source as well.
The SimpleScalar architecture, like the MIPS architecture [4],
supports both big-endian and little-endian executables. The tool
set supports compilation for either of these targets; the names for
the big-endian and little-endian architecture are ssbig-na-sstrix
and sslittle-na-sstrix, respectively. You should use the target
endian-ness that matches your host platform; the simulators may
not work correctly if you force the compiler to provide crossendian support. To determine which endian your host uses, run
the endian program located in the simplesim-2.0/ directory. For simplicity, the following instructions will assume a bigendian installation. In the following instructions, we will refer to
the directory in which you are installing SimpleScalar as
$IDIR/.
The simulators come equipped with their own loader, and
thus you do not need to build the GNU binary utilities to run simulations. However, many of these utilities are useful, and we recommend that you install them. If desired, build the GNU binary
utilities1:

sstrix --with-gnu-as --with-gnu-ld --prefix=$IDIR

make
make install

$HOST here is a canonical configuration string that represents

your host architecture and system (CPU-COMPANY-SYSTEM).
The string for a Sparcstation running SunOS would be sparc-sunsunos4.1.3, running Solaris: sparc-sun-solaris2, a 386 running
Solaris: i386-sun-solaris2.4, etc. A complete list of supported
$HOST strings resides in $IDIR/gcc-2.6.3/INSTALL.
This installation will create the needed directories in $IDIR
(these include bin/, lib/, include/, and man/). Once the
binutils have been built, build the simulators themselves. This is
necessary to do before building GCC, since one of the binaries is
needed for the cross-compiler build. You should edit $IDIR/
simplesim-2.0/Makefile to use the desired compile flags
(e.g., the correct optimization level). To use the GNU BFD
loader instead of the custom loader in the simulators, uncomment
-DBFD_LOADER in the Makefile. To build the simulators:
cd $IDIR/simplesim-2.0
make

If desired, build the compiler:

cd $IDIR/gcc-2.6.3
configure --host=$HOST --target=ssbig-nasstrix --with-gnu-as --with-gnu-ld --prefix=$IDIR
make LANGUAGES=c
../simplesim-2.0/sim-safe ./enquire -f >!
float.h-cross
make install

We provide pre-built copies of the necessary libraries in ssbigna-sstrix/lib/, so you do not need to build the code in
glibc-1.09, unless you change the library code. Building these
libraries is tricky, and we do not recommend it unless you have a
specific need to do so. In that event, to build the libraries:

cd $IDIR/binutils-2.5.2
configure --host=$HOST --target=ssbig-na-

1. You must have GNU Make to do the majority of installations described

in this document. To check if you have the GNU version, execute make v or gmake -v. The GNU version understands this switch and displays
version information.

cd $IDIR/glibc-1.09
configure --prefix=$IDIR/ssbig-na-sstrix
ssbig-na-sstrix

description of each. Both the number and the semantics of the

registers are identical to those in the MIPS-IV ISA.
In Figure 3, we depict the three instruction encodings of SimpleScalar instructions: register, immediate, and jump formats. All
instructions are 64 bits in length.
The register format is used for computational instructions.
The immediate format supports the inclusion of a 16-bit constant.
The jump format supports specification of 24-bit jump targets.
The register fields are all 8 bits, to support extension of the architected registers to 256 integer and floating point registers. Each
instruction format has a fixed-location, 16-bit opcode field that
facilitates fast instruction decoding.
The annote field is a 16-bit field that can be modified postcompile, with annotations to instructions in the assembly files.
The annotation interface is useful for synthesizing new instructions without having to change and recompile the assembler.
Annotations are attached to the opcode, and come in two flavors:
bit and field annotations. A bit annotation is written as follows:

setenv CC $IDIR/bin/ssbig-na-sstrix-gcc
unsetenv TZ
unsetenv MACHINE
make
make install

Note that you must have already built the SimpleScalar simulators to build this library, since the glibc build requires a compiled
simulator to test target machine-specific parameters such as
endian-ness.
If you have FORTRAN benchmarks, you will need to build
f2c:
cd $IDIR/f2c-1994.09.27
make
make install

The entire tool set should now be ready for use. We provide precompiled test binaries (big- and little-endian) and their sources in
$IDIR/simplesim2.0/tests). To run a test:
cd $IDIR/simplesim-2.0
sim-safe tests/bin.big/test-math

lw/a

$r6,4($r7)

The annotation in this example is /a. It specifies that the first bit
of the annotation field should be set. Bit annotations /a through /p
set bits 0 through 15, respectively. Field annotations are written
in the form:

The test should generate about a page of output, and will run very
quickly. The release has been ported toand should run onthe
following systems:
- gcc/AIX 413/RS6000
- xlc/AIX 413/RS6000
- gcc/HPUX/PA-RISC
- gcc/SunOS 4.1.3/SPARC
- gcc/Linux 1.3/x86
- gcc/Solaris 2/SPARC
- gcc/Solaris 2/x86
- gcc/DEC Unix 3.2/Alpha
- c89/DEC Unix 3.2/Alpha
- gcc/FreeBSD 2.2/x86
- gcc/WindowsNT/x86

lw/6:4(7)

$r6,4($r7)

This annotation sets the specified 3-bit field (from bit 4 to bit 6
within the 16-bit annotation field) to the value 7.
System calls in SimpleScalar are managed by a proxy handler
(located in syscall.c) that intercepts system calls made by
the simulated binary, decodes the system call, copies the system
call arguments, makes the corresponding call to the hosts operating system, and then copies the results of the call into the simulated programs memory. If you are porting SimpleScalar to a
new platform, you will have to code the system call translation
from SimpleScalar to your host machine in syscall.c. A list
of all SimpleScalar system calls is provided in Appendix B.
SimpleScalar uses a 31-bit address space, and its virtual
memory is laid out as follows:

3 The Simplescalar architecture

The SimpleScalar architecture is derived from the MIPS-IV
ISA [4]. The tool suite defines both little-endian and big-endian
versions of the architecture to improve portability (the version
used on a given host machine is the one that matches the endianness of the host). The semantics of the SimpleScalar ISA are a
superset of MIPS with the following notable differences and
additions:

There are no architected delay slots: loads, stores, and control transfers do not execute the succeeding instruction.

Loads and stores support two addressing modesfor all

data typesin addition to those found in the MIPS architecture. These are: indexed (register+register), and auto-increment/decrement.

A square-root instruction, which implements both singleand double-precision floating point square roots.

An extended 64-bit instruction encoding.

We list all SimpleScalar instructions in Figure 2. We provide
a complete list of the instruction semantics (as implemented in
the simulator) in Appendix A. In Table 1, we list the architected
registers in the SimpleScalar architecture, their hardware and
software names (which are recognized by the assembler), and a

0x00000000
0x00400000
0x10000000
0x7fffc000

Unused
Start of text segment
Start of data segment
Stack base (grows down)

The top of the data segment (which includes init and bss) is held
in mem_brk_point. The areas below the text segment and
above the stack base are unused.

4 Simulator internals
In this section, we describe the functionality of the processor
simulators that accompany the tool set. We describe each of the
simulators, their functionality, command-line arguments, and
internal structures.
The compiler outputs binaries that are compatible with the
MIPS ECOFF object format. Library calls are handled with the
ported version of GNU GLIBC and POSIX-compliant Unix system calls. The simulators currently execute only user-level code.
All SimpleScalar-related extensions to GCC are contained in the
config/ss subdirectory of the GCC source tree that comes

Control

Load/Store

Integer Arithmetic

Floating Point Arithmetic

j - jump
jal - jump and link
jr - jump register
jalr - jump and link register
beq - branch == 0
bne - branch != 0
blez - branch <= 0
bgtz - branch > 0
bltz - branch < 0
bgez - branch >= 0
bct - branch FCC TRUE
bcf - branch FCC FALSE

lb - load byte
lbu - load byte unsigned
lh - load half (short)
lhu - load half (short) unsigned
lw - load word
dlw - load double word
l.s - load single-precision FP
l.d - load double-precision FP
sb - store byte
sbu - store byte unsigned
sh - store half (short)
shu - store half (short) unsigned
sw - store word
dsw - store double word
s.s - store single-precision FP
s.d - store double-precision FP

add - integer add

addu - integer add unsigned
sub - integer subtract
subu - integer subtract unsigned
mult - integer multiply
multu - integer multiply unsigned
div - integer divide
divu - integer divide unsigned
and - logical AND
or - logical OR
xor - logical XOR
nor - logical NOR
sll - shift left logical
srl - shift right logical
sra - shift right arithmetic
slt - set less than
sltu - set less than unsigned

add.s - single-precision (SP) add

add.d - double-precision (DP) add
sub.s - SP subtract
sub.d - DP subtract
mult.s - SP multiply
mult.d - DP multiply
div.s - SP divide
div.d - DP divide
abs.s - SP absolute value
abs.d - DP absolute value
neg.s - SP negation
neg.d - DP negation
sqrt.s - SP square root
sqrt.d - DP square root
cvt - int., single, double conversion
c.s - SP compare
c.d - DP compare

addressing modes:
(C)
(reg+C) (with pre/post inc/dec)
(reg+reg) (with pre/post inc/dec)

Miscellaneous
nop - no operation
syscall - system call
break - declare program error

Figure 2. Summary of SimpleScalar instructions

Hardware Name
$0
$1
$2-$3
$4-$7
$8-$15
$16-$23
$25-$25
$26-$27
$28
$29
$30
$31
$hi
$lo
$f0-$f31
$fcc

Software Name
$zero
$at
$v0-$v1
$a0-$a3
$t0-$t7
$s0-$s7
$t8-$t9
$k0-$k1
$gp
$sp
$s8
$ra
$hi
$lo
$f0-$f31
$fcc

Description
zero-valued source/sink
reserved by assembler
fn return result regs
fn argument value regs
temp regs, caller saved
saved regs, callee saved
temp regs, caller saved
reserved by OS
global pointer
stack pointer
saved regs, callee saved
return address reg
high result register
low result register
floating point registers
floating point condition code

Table 1: SimpleScalar architecture register definitions

16-annote

16-opcode

8-rs

8-rt

8-rd

8-ru/shamt

Register format:
63

32 31
16-annote

16-opcode

0
8-rs

16-imm

8-rt

Immediate format:
63

32 31
16-annote

16-opcode

6-unused

0
26-target

Jump format:
63

32 31

Figure 3. SimpleScalar architecture instruction formats

time is not needed.

sim-cache accepts the following arguments, in addition to the
universal arguments described in Section 4:

with the distribution.

The architecture is defined in ss.def, which contains a
macro definition for each instruction in the instruction set. Each
macro defines the opcode, name, flags, operand sources and destinations, and actions to be taken for a particular instruction.
The instruction actions (which appear as macros) that are
common to all simulators are defined in ss.h. Those actions
that require different implementations in different simulators are
defined in each simulator code file.
When running a simulator, main() (defined in main.c)
does all the initialization and loads the target binary into memory. The routine then calls sim_main(), which is simulatorspecific, defined in each simulator code file. sim_main() predecodes the entire text segment for faster simulation, and then
begins simulation from the target program entry point.
The following command-line arguments are available in all
simulators included with the release:
-h
prints the simulator help message.
-d
turn on the debug message.
-i
start execution in the DLite! debugger (see
Section 5.2). This option is not supported in
the sim-fast simulator.
-q
terminate immediately (for use with -dumpconfig).
-dumpconfig <file>
generate a configuration file saving the command-line parameters. Comments are permitted in the config files, and begin with a #.
-config <file> read in and use a configuration file. These
files may reference other config files.

-cache:dl1 <config>
configures a level-one data cache.
-cache:dl2 <config>
configures a level-two data cache.
-cache:il1 <config>
configures a level-one instr. cache.
-cache:il2 <config>
configures a level-two instr. cache.
-tlb:dtlb <config>
configures the data TLB.
-tlb:itlb <config>
configures the instruction TLB.
-flush <boolean>
flush all caches on a system call;
(<boolean> = 0 | 1 | true | TRUE | false | FALSE).
-icompress
remap
SimpleScalars
64-bit
instructions to a 32-bit equivalent in
the simulation (i.e., model a
machine with 4-word instructions).
-pcstat <stat>
generate a text-based profile, as
described in Section 4.3.
The cache configuration (<config>) is formatted as follows:
<name>:<nsets>:<bsize>:<assoc>:<repl>

Each of these fields has the following meaning:

<name>
cache name, must be unique.
<nsets>
number of sets in the cache.
<bsize>
block size (for TLBs, use the page size).
<assoc>
associativity of the cache (power of two).
<repl>
replacement policy (l | f | r), where
l = LRU, f = FIFO, r = random replacement.
The cache size is therefore the product of <nsets>, <bsize>, and
<assoc>. To have a unified level in the hierarchy, point the
instruction cache to the name of the data cache in the corresponding level, as in the following example:

4.1 Functional simulation

The fastest, least detailed simulator (sim-fast) resides in
sim-fast.c. sim-fast does no time accounting, only functional simulationit executes each instruction serially, simulating no instructions in parallel. sim-fast is optimized for raw
speed, and assumes no cache, instruction checking, and has no
support for DLite!.
A separate version of sim-fast, called sim-safe, also performs
functional simulation, but checks for correct alignment and
access permissions for each memory reference. Although similar,
sim-fast and sim-safe are split (i.e., protection is not toggled
with a command-line argument in a merged simulator) to maximize performance. Neither of the simulators accept any additional command-line arguments. Both versions are very simple:
less than 300 lines of codethey therefore make good starting
points for understanding the internal workings of the simulators.
In addition to the simulator file, both sim-fast and sim-safe use
the following code files (not including header files): main.c,
syscall.c, memory.c, regs.c, loader.c, ss.c,
endian.c, and misc.c. sim-safe also uses dlite.c.

-cache:il1
-cache:il2
-cache:dl1
-cache:dl2

il1:128:64:1:l
dl2
dl1:256:32:1:l
ul2:1024:64:2:l

The defaults used in sim-cache are as follows:

L1 instruction cache:
L1 data cache:
L2 unified cache:
instruction TLB:
data TLB:

il1:256:32:1:l
dl1:256:32:1:l
ul2:1024:64:4:l
itlb:16:4096:4:l
dtlb:32:4096:4:l

(8 KB)
(8 KB)
(256 KB)
(64 entries)
(128 entries)

sim-cheetah is based on work performed by Ragin Sugumar and

Santosh Abraham while they were at the University of Michigan.
It uses their Cheetah cache simulation engine [6] to generate simulation results for multiple cache configurations with a single
simulation. The Cheetah engine simulates fully associative
caches efficiently, as well as simulating a sometimes-optimal
replacement policy. This policy was called MIN by Belady [1],
although the simulator refers to it as opt. Opt uses future knowledge to select a replacement; it chooses the block that will be referenced the furthest in the future (if at all). This policy is optimal
for read-only instruction streams. It is not optimal for write-back
caches because it may be more expensive to replace a block referenced further in the future if the block must be written back, as
opposed to a clean block referenced slightly less far in the future.

4.2 Cache simulation

The SimpleScalar distribution comes with two functional
cache simulators; sim-cache and sim-cheetah. Both use the file
cache.c, and they use sim-cache.c and sim-cheetah.c, respectively. These simulators are ideal for fast simulation of caches if the effect of cache performance on execution

-pcstat <stat>

Horwitz et al. [3] formally described an optimal algorithm that

includes writes; however, only MIN is implemented in the simulator.
We have included the Cheetah engine as a stand-alone library,
which is built and resides in the libcheetah/ directory. simcheetah accepts the following command-line arguments, in addition to those listed at the beginning of Section 4:

where <stat> is the integer counter that you

wish to profile by text address.
To generate the statistics for the profile, follow the following
example:
sim-profile -pcstat sim_num_insn test-math >&!
test-math.out
objdump -dl test-math >! test-math.dis
textprof.pl test-math.dis test-math.out
sim_num_insn_by_pc

-refs [inst | data | unified]

specify which reference stream to analyze.
-C [fa | sa | dm]

-R [lru | opt]
-a <sets>
-b <sets>
-l <line>
-n <assoc>
-in <interval>
-M <size>
-C <size>

We show a segment of the text profile output in Figure 4. Make

sure that objdump is the version created when compiling the
binutils. Also, the first line of textprof.pl must be changed
to reflect your systems path to Perl (which must be installed on
your system for you to use this script). As an aside, note that taddrprof is equivalent to -pcstat sim_num_insn.

fully associative, set associative, or directmapped cache.

replacement policy.
log base 2 minimum bound on number of
sets to simulate simultaneously.
log base 2 maximum bound on set number.
cache line size (in bytes).
maximum associativity to analyze (in log
base 2).
cache size interval to report when simulating
fully associative caches.
maximum cache size of interest.
cache size for direct-mapped analyses.

4.4 Out-of-order processor timing simulation

The most complicated and detailed simulator in the distribution, by far, is sim-outorder (the main code file for which is
sim-outorder.cabout 3500 lines long). This simulator
supports out-of-order issue and execution, based on the Register
Update Unit [5]. The RUU scheme uses a reorder buffer to automatically rename registers and hold the results of pending
instructions. Each cycle the reorder buffer retires completed
instructions in program order to the architected register file.
The processors memory system employs a load/store queue.
Store values are placed in the queue if the store is speculative.
Loads are dispatched to the memory system when the addresses
of all previous stores are known. Loads may be satisfied either by
the memory system or by an earlier store value residing in the
queue, if their addresses match. Speculative loads may generate
cache misses, but speculative TLB misses stall the pipeline until
the branch condition is known.
We depict the simulated pipeline of sim-outorder in
Figure 5. The main loop of the simulator, located in
sim_main(), is structured as follows:

Both of these simulators are ideal for performing high-level

cache studies that do not take access time of the caches into
account (e.g., studies that are concerned only with miss rates). To
measure the effect of cache organization upon the execution time
of real programs, however, the timing simulator described in
Section 4.4 must be used.

4.3 Profiling
The distribution comes with a functional simulator that produces voluminous and varied profile information. sim-profile
can generate detailed profiles on instruction classes and
addresses, text symbols, memory accesses, branches, and data
segment symbols.
sim-profile takes the following command-line arguments,
which toggle the various profiling features:
-iclass
instruction class profiling (e.g. ALU,
branch).
-iprof
instruction profiling (e.g., bnez, addi).
-brprof
branch class profiling (e.g., direct, calls, conditional).
-amprof
addr. mode profiling (e.g., displaced, R+R).
-segprof
load/store segment profiling (e.g., data,
heap).
-tsymprof
execution profile by text symbol (functions).
-dsymprof
reference profile by data segment symbol.
-taddrprof
execution profile by text address.
-all
turn on all profiling listed above.
Three of the simulators (sim-profile, sim-cache, and sim-outorder) support text segment profiles for statistical integer
counters. The supported counters include any added by users, so
long as they are correctly registered with the SimpleScalar
stats package included with the simulator code (see Section 4.5).
To use the counter profiles, simply add the command-line flag:

ruu_init();
for (;;) {
ruu_commit();
ruu_writeback();
lsq_refresh();
ruu_issue();
ruu_dispatch();
ruu_fetch();
}

This loop is executed once for each target (simulated)

machine cycle. By walking the pipeline in reverse, inter-stage
latch synchronization can be handled correctly with only one
pass through each stage. When the target program terminates
with an exit() system call, the simulator performs a
longjmp() to main() to generate the statistics.
The fetch stage of the pipeline is implemented in
ruu_fetch(). The fetch unit models the machine instruction
bandwidth, and takes the following inputs: the program counter,
the predictor state, and misprediction detection from the branch
execution unit(s). Each cycle, it fetches instructions from only
one I-cache line (and it blocks on an I-cache miss until the miss

executed
13 times

never
executed

{
{

00401a10: (
strtod.c:79
00401a18: (
strtod.c:87
00401a20:
00401a28:
strtod.c:89
00401a30: (
00401a38: (
00401a40: (

13,

0.01): <strtod+220> addiu $a1[5],$zero[0],1

13,

0.01): <strtod+228> bc1f 00401a30 <strtod+240>

: <strtod+230> addiu $s1[17],$s1[17],1
: <strtod+238> j 00401a58 <strtod+268>

13,
13,
13,

0.01): <strtod+240> mul.d $f2,$f20,$f4

0.01): <strtod+248> addiu $v0[2],$v1[3],-48
0.01): <strtod+250> mtc1 $v0[2],$f0

Figure 4. Sample output from text segment statistical profile

Fetch

Dispatch

Scheduler

Exec

Memory
scheduler

Mem

I-Cache

Writeback

Commit

D-Cache D-TLB
Virtual memory
Figure 5. Pipeline for sim-outorder
load is sent to the memory system.
The execute stage is also handled in ruu_issue(). Each
cycle, the routine gets as many ready instructions as possible
from the scheduler queue (up to the issue width). The functional
units availability is also checked, and if they have available
access ports, the instructions are issued. Finally, the routine
schedules writeback events using the latency of the functional
units (memory operations probe the data cache to obtain the correct latency of the operation). Data TLB misses stall the issue of
the memory operation, are serviced in the commit stage of the
pipeline, and currently assume a fixed latency. The functional
units latencies are hardcoded in the definition of
fu_config[] in sim-outorder.c.
The writeback stage resides in ruu_writeback(). Each
cycle it scans the event queue for instruction completions. When
it finds a completed instruction, it walks the dependence chain of
instruction outputs to mark instructions that are dependent on the
completed instruction. If a dependent instruction is waiting only
for that completion, the routine marks it as ready to be issued.
The writeback stage also detects branch mispredictions; when it
determines that a branch misprediction has occurred, it rolls the
state back to the checkpoint, discarding the erroneously issued
instructions.
ruu_commit() handles the instructions from the writeback
stage that are ready to commit. This routine does in-order committing of instructions, updating of the data caches (or memory)
with store values, and data TLB miss handling. The routine keeps

completes). After fetching the instructions, it places them in the

dispatch queue, and probes the line predictor to obtain the correct
cache line to access in the next cycle.
The code for the dispatch stage of the pipeline resides in
ruu_dispatch(). This routine is where instruction decoding
and register renaming is performed. The function uses the
instructions in the input queue filled by the fetch stage, a pointer
to the active RUU, and the rename table. Once per cycle, the dispatcher takes as many instructions as possible (up to the dispatch
width of the target machine) from the fetch queue and places
them in the scheduler queue. This routine is the one in which
branch mispredictions are noted. (When a misprediction occurs,
the simulator uses speculative state buffers, which are managed
with a copy-on-write policy). The dispatch routine enters and
links instructions into the RUU and the load/store queue (LSQ),
as well as splitting memory operations into two separate instructions (the addition to compute the effective address and the memory operation itself).
The issue stage of the pipeline is contained in
ruu_issue() and lsq_refresh(). These routines model
instruction wakeup and issue to the functional units, tracking register and memory dependences. Each cycle, the scheduling routines locate the instructions for which the register inputs are all
ready. The issue of ready loads is stalled if there is an earlier
store with an unresolved effective address in the load/store
queue. If the address of the earlier store matches that of the waiting load, the store value is forwarded to the load. Otherwise, the

order, with the following additions:

-cache:dl1lat <cycles>
specify the hit latency of the L1 data cache.
The default is 1 cycle.
-cache:d12lat <cycles>
specify the hit latency of the L2 data cache.
The default is 6 cycles.
-cache:il1lat <cycles>
specify the hit latency of the L1 instruction
cache. The default is 1 cycle.
-cache:il2lat <cycles>
specify the hit latency of the L2 instruction
cache. The default is 6 cycles.
-mem:lat <1st> <next>
specify main memory access latency (first,
rest). The defaults are 18 cycles and 2 cycles.
-mem:width <bytes>
specify width of memory bus in bytes. The
default is 8 bytes.
-tlb:lat <cycles>
specify latency (in cycles) to service a TLB
miss. The default is 30 cycles.

retiring instructions at the head of the RUU that are ready to

commit until the head instruction is one that is not ready. When
an instruction is committed, its result is placed into the architected register file, and the RUU/LSQ resources devoted to that
instruction are reclaimed.
sim-outorder runs about an order of magnitude slower than
sim-fast. In addition to the arguments listed at the beginning of
Section 4, sim-outorder uses the following command-line arguments:
Specifying the processor core
-fetch:ifqsize <size>
set the fetch width to be <size> instructions.
Must be a power of two. The default is 4.
-fetch:speed <ratio>
set the ratio of the front end speed relative to
the execution core (allowing <ratio> times as
many instructions to be fetched as decoded
per cycle).
-fetch:mplat <cycles>
set the branch misprediction latency. The
default is 3 cycles.
-decode:width <insts>
set the decode width to be <insts>, which
must be a power of two. The default is 4.
-issue:width <insts>
set the maximum issue width in a given
cycle. Must be a power of two. The default is
4.
-issue:inorder force the simulator to use in-order issue. The
default is false.
-issue:wrongpath
allow instructions to issue after a misspeculation. The default is true.
-ruu:size <insts>
capacity of the RUU (in instructions). The
default is 16.
-lsq:size <insts>
capacity of the load/store queue (in instructions). The default is 8.
-res:ialu <num>
specify number of integer ALUs. The default
is 4.
-res:imult <num>
specify number of integer multipliers/dividers. The default is 1.
-res:memports <num>
specify number of L1 cache ports. The
default is 2.
-res:fpalu <num>
specify number of floating point ALUs. The
default is 4.
-res: fpmult <num>
specify number of floating point multipliers/
dividers. The default is 1.

Specifying the branch predictor

Branch prediction is specified by choosing the following flag
with one of the six subsequent arguments. The default is a bimodal predictor with 2048 entries.
-bpred <type>
nottaken
always predict not taken.
taken
always predict taken.
perfect
perfect predictor.
bimod
bimodal predictor, using a branch target
buffer (BTB) with 2-bit counters.
2lev
2-level adaptive predictor.
comb
combined predictor (bimodal and 2-level
adaptive).
The predictor-specific arguments are listed below:
-bpred:bimod <size>
set the bimodal predictor table size to be
<size> entries.
-bpred:2lev <l1size> <l2size> <hist_size> <xor>
specify the 2-level adaptive predictor.
<l1size> specifies the number of entries in
the first-level table, <l2size> specifies the
number of entries in the second-level table,
<hist_size> specifies the history width, and
<xor> allows you to xor the history and the
address in the second level of the predictor.
This organization is depicted in Figure 6. In
Table 2 we show how these parameters correspond to modern prediction schemes. The
default settings for the four parameters are 1,
1024, 8, and 0, respectively.
-bpred:comb <size>
set the meta-table size of the combined predictor to be <size> entries. The default is
1024.

Specifying the memory hierarchy

All of the cache arguments and formats used in sim-cache
(listed at the beginning of Section 4.2) are also used in sim-out-

pattern
history

2-bit
predictors

branch
address

branch
prediction
l2size

l1size

hist_size

Figure 6. 2-level adaptive predictor structure

predictor

l1_size

hist_size

l2_size

xor

GAg

2W

GAp

>2W

PAg

2W

PAp

2N+W

gshare

2W

Table 2: Branch predictor parameters

-bpred:ras <size>
set the return stack size to <size> (0 entries
means to return stack). The default is 8.
entries.
-bpred:btb <sets> <assoc>
configure the BTB to have <sets> sets and an
associativity of <assoc>. The defaults are
512 sets and an associativity of 4.
-bpred:spec_update <stage>
allow speculative updates of the branch predictor in the decode or writeback stages
(<stage> = [ID|WB]). The default is nonspeculative updates in the commit stage.

Visualization
-pcstat <stat>

record statistic <stat> by text address;

described in Section 4.3.
-ptrace <file> <range>
pipeline tracing, described in Section 5.

4.5 Simulator code file descriptions

The following list describes the functionality of the C code
files in the simplesim-2.0/ directory, which are used by all
of the simulators.

bitmap.h: Contains support macros for performing bitmap manipulation.

bpred.[c,h]: Handles the creation, functionality, and

updates of the branch predictors. bpred_create(),
bpred_lookup(), and bpred_update() are the key
interface functions.

cache.[c,h]: Contains general functions to support

10

multiple cache types (e.g., TLBs, instruction and data

caches). Uses a linked-list for tag comparisons in caches of
low associativity (less than or equal to four), and a hash
table for tag comparisons in higher-associativity caches.
The important interfaces are cache_create(),
cache_access(), cache_probe(),
cache_flush(), and cache_flush_addr().
dlite.[c,h]: Contains the code for DLite!, the sourcelevel target program debugger.
endian.[c,h]: Defines a few simple functions to determine byte- and word-order on the host and target platforms.
eval.[c,h]: Contains code to evaluate expressions, used
in DLite!.
eventq.[c,h]: Defines functions and macros to handle
ordered event queues (used for ordering writebacks). The
important interface functions are eventq_queue() and
eventq_service_events().
loader.[c,h]: Loads the target program into memory,
sets up the segment sizes and addresses, sets up the initial
call stack, and obtains the target program entry point. The
interface is ld_load_prog().
main.c: Performs all initialization and launches the main
simulator function. The key functions are
sim_options(), sim_config(), sim_main(),
and sim_stats().
memory.[c,h]: Contains functions for reading from,
writing to, initializing, and dumping the contents of the target main memory. Memory is implemented as a large flat
space, each portion of which is allocated on demand.
mem_access() is the important interface function.
misc.[c,h]: Contains numerous useful support functions, such as fatal(), panic(), warn(), info(),
debug(), getcore(), and elapsed_time().
options.[c,h]: Contains the SimpleScalar options
package code, used to process command-line arguments
and/or option specifications from config files. Options are
registered with an option database (see the functions called
opt_reg_*()). opt_print_help() generates a help
listing, and opt_print_options() prints the current
options state.
ptrace.[c,h]: Contains code to collect and produce
pipeline traces from sim-outorder.
range.[c,h]: Holds code that interprets program range
commands used in DLite!.
regs.[c,h]: Contains functions to initialize the register
files and dump their contents.
resource.[c,h]: Contains code to manage functional
unit resources, divided up into classes. The three defined
functions create the resource pools and busy tables
(res_create_pool()), return a resource from the specified pool if available (reg_get()), and dump the contents of a pool (res_dump()).
sim.h: Contains a few extern variable declarations and
function prototypes.
stats.[c,h]: Contains routines to handle statistics measuring target program behavior. As with the options pack-

The traces may be viewed with the pipeview.pl Perl script,

which is provided in the simplesim-2.0 directory. (You will have
to update the first line of pipeview.pl to have the correct path
to your local Perl binary, and you must have Perl installed on
your system).

age, counters are registered by type with an internal

database. The stat_reg_*() routines register counters
of various types, and stat_reg_formula() allows you
to register expressions constructed of other statistics.
stat_print_stats() prints all registered statistics.
The statistics package also has facilities to measure distributions; stat_reg_dist() creates an array distribution,
stat_reg_sdist() creates a sparse array distribution,
and stat_add_sample() updates a distribution.
ss.[c,h]: Defines macros to expedite the processing of
instructions, numerous constants needed across simulators,
and a function to print out individual instructions in a readable format.
ss.def: Holds a list of macro calls (the macros are defined
in the simulators and ss.h and ss.c), each of which
defines an instruction. The macro calls accept as arguments
the opcode, name of the instruction, sources, destinations,
actions to execute, and other information. This file serves as
the definition of the instruction set.
symbol.[c,h]: Holds routines to handle program symbol and line information (used in DLite!).
syscall.[c,h]: Contains code that acts as the interface
between the SimpleScalar system calls (which are POSIXcompliant) and the system calls on the host machine.
sysprobe.c: Determines byte and word order on the host
platform, and generates appropriate compiler flags.
version.h: Defines the version number and release date
of the distribution.

pipeview.pl <ptrace_file>

We depict sample output from the pipetracer in Figure 7.

5.2 The DLite! debugger

Release 2.0 of SimpleScalar includes a lightweight symbolic
debugger called DLite!, which runs with all simulators except for
sim-fast. DLite! allows you to step through the benchmark target
code, not the simulator code. The debugger can be incorporated
into a simulator by adding only four function calls (which have
already been added to all simulators in the distribution). The
needed four function prototypes are in dlite.h.
To use the debugger in a simulation, add the -i option
(which stands for interactive) to the simulator command line.
Below we list the set of commands that DLite! accepts.
Getting help and getting out:
help [string]
print command reference.
version
print DLite! version information.
quit
exit simulator.
terminate
generate statistics and exit simulator.
Running and setting breakpoints:
step
execute next instruction and break.
cont [addr]
continue execution (optionally continuing
starting at <addr>).
break <addr> set breakpoint at <addr>, returns <id> of
breakpoint.
dbreak <addr> [r,w,x]
set data breakpoint at <addr> for (r)ead,
(w)rite, and/or e(x)ecute, returns <id> of
breakpoint.
rbreak <range> [r,w,x]
set breakpoint at <range> for (r)ead, (w)rite,
and/or e(x)ecute, returns <id> of breakpoint.
breaks
list active code and data breakpoints.
delete <id>
delete breakpoint <id>.
clear
clear all breakpoints (code and data).

5 Utilities
In this section we describe the utilities that accompany the
SimpleScalar tool set; pipeline tracing and a source-level debugger.

5.1 Out-of-order pipeline tracing

The tool set provides the ability to extract and view traces of
the out-of-order pipeline. Using the -ptrace option, a detailed
history of all instructions executed in a range may be saved to a
file. The information saved includes instruction fetch, retirement,
and stage transitions. The syntax of this command is as follows:
-ptrace <file> <start>:<end>
<file> is the file to which the trace will be
saved. <start> and <end> are the instruction
numbers at which the trace will be started
and stopped. If they are left blank, the trace
will start at the beginning and/or stop at the
end of the program, respectively.
For example:
-ptrace FOO.trc 100:500
trace from instructions 100 to 500, store the
trace in file FOO.src.
-ptrace FOO.trc :10000
trace from program beginning to instruction
10000.
-ptrace FOO.trc :
trace the entire program execution.

Printing information:
print [modifiers] <expr>
print the value of <expr> using optional
modifiers.
display [modifiers] <expr>
display the value of <expr> using optional
modifiers.
option <string> print the value of option <string>.
options
print the values of all options.
stat <string> print the value of a statistical variable.
stats
print the values of all statistical variables.
whatis <expr> print the type of <expr>.
regs
print all register contents.
iregs
print all instruction register contents.

11

new cycle
indicator

@ 610

new instruction
definitions

gf = 0x0040d098: addiu
gg = 0x0040d0a0: beq

current pipeline
state

[IF]
gf
gg

[DA]
gb
gc
gd\
ge

[EX]
fy
fz
ga+

[WB]
fr\
fs
ft
fu

[CT]
fq

inst. being
fetched, or in
fetch queue

inst. being
decoded, or
awaiting issue

inst.
executing

inst. writing
results into
RUU, or
awaiting retire

inst. retiring
results to
register file

pipeline event:
(misprediction
detected), see output
header for event defs

r2, r4, -1
r3, r5, 0x30

Figure 7. Example of sim-outorder pipetrace

w
t
o
x
1
f
d
c
s

fpregs
print all floating point register contents.
mstate [string] print machine-specific state.
dump <addr> [count]
dump memory at <addr> (optionally for
<count> words).
dis <addr> [count]
disassemble instructions at <addr> (optionally for <count> instructions).
symbols
print the value of all program symbols.
tsymbols
print the value of all program text symbols.
dsymbols
print the value of all program data symbols.
symbol <string>
print the value of symbol <string>.

print a word (default)

print in decimal format (default)
print in octal format
print in hex format
print in binary format
print float
print double
print character
print string

Examples of legal commands:

break main+8
break 0x400148
dbreak stdin w
dbreak sys_count wr
rbreak @main:+279
rbreak 2000:3500
rbreak #:100
cycle 0 to cycle 100
rbreak :
entire execution

Legal arguments:
Arguments <addr>, <cnt>, <expr>, and <id> are any legal
expression:
<expr> <factor> +|- <expr>
<factor> <term> *|/ <factor>
<term> ( <expr> )
| - <term> | <const> | <symbol> | <file:loc>
<symbol> <literal> | <function name> | <register>
<literal> [0-9]+ | 0x[0-9,a-f]+ | 0[0-7]+
<register> $r[0-31] | $f[0-31] | $pc | $fcc | $hi | $lo

6 Summary
The SimpleScalar tool set was written by Todd Austin over
about one and a half years, between 1994 and 1996. He continues
to add improvements and updates. The ancestors of the tool set
date back to the mid to late 1980s, to tools written by Manoj
Franklin. At the time the tools were developed, both individuals
were research assistants at the University of Wisconsin-Madison
Computer Sciences Department, supervised by Professor Guri
Sohi. Scott Breach provided valuable assistance with the implementation of the proxy system calls. The first release was assembled, debugged, and documented by Doug Burger, also a
research assistant at Wisconsin, who is the maintainer of the second release as well. Kevin Skadron, currently at Princeton,
implemented many of the more recent branch prediction mechanisms.
Many exciting extensions to SimpleScalar are both underway
and planned. Efforts have begun to extend the processor simula-

Legal ranges:
<range> <address> | <instruction> | <cycle>
<address> @<function name>:{+<literal>}
<instruction> {<literal>}:{<literal>}
<cycle> #{<literal>}:{<literal>}
Omitting optional arguments to the left of the colon will default
to the smallest value permitted in that range. Omitting an
optional argument at the right of the colon will default to the
largest value permitted in that range.
Legal command modifiers:
b print a byte
h print a half (short)

12

Semantics:

tors to simulate multithreaded processors and multiprocessors. A

Linux port to SimpleScalar (enabling simulation of the OS on a
kernel with publicly available sources) is planned, using devicelevel emulation and a user-level file system. Other plans include
extending the tool set to simulate ISAs other than SimpleScalar
and MIPS (Alpha and SPARC ISA support will be the first additions).
As they stand now, these tools provide researchers with a simulation infrastructure that is fast, flexible, and efficient. Changes
in both the target hardware and software may be made with minimal effort. We hope that you find these tools useful, and encourage you to contact us with ways that we can improve the release,
documentation, and the tools themselves.

SET_NPC((CPC\&0xf0000000) | (TARGET<<2))
SET_GPR(31, CPC + 8))

JR:
Opcode:
Format:
Semantics:

Jump to register address.

0x03
JR rs

JALR:
Opcode:
Format:
Semantics:

Jump to register address and link.

0x04
JALR rs

TALIGN(GPR(RS))
SET_NPC(GPR(RS))

TALIGN(GPR(RS))

SET_GPR(RD, CPC + 8)
SET_NPC(GPR(RS))

References
[1]

[2]

[3]

[4]
[5]

[6]

L. A. Belady. A Study of Replacement Algorithms for a

Virtual-Storage Computer. IBM Systems Journal, 5(2):78
101, 1966.
Doug Burger, Todd M. Austin, and Steven Bennett. Evaluating Future Microprocessors: the SimpleScalar Tool Set.
Technical Report 1308, Computer Sciences Department,
University of Wisconsin, Madison, WI, July 1996.
L. P. Horwitz, R. M. Karp, R. E. Miller, and A. Winograd.
Index Register Allocation. Journal of the ACM, 13(1):43
61, January 1966.
Charles Price. MIPS IV Instruction Set, revision 3.1. MIPS
Technologies, Inc., Mountain View, CA, January 1995.
Gurindar S. Sohi. Instruction Issue Logic for High-Performance, Interruptible, Multiple Functional Unit, Pipelined
Computers. IEEE Transactions on Computers, 39(3):349
359, March 1990.
Rabin A. Sugumar and Santosh G. Abraham. Efficient
Simulation of Caches under Optimal Replacement with
Applications to Miss Characterization. In Proceedings of
the 1993 ACM Sigmetrics Conference on Measurements
and Modeling of Computer Systems, pages 2435, May
1993.

A Instruction set definition

This appendix lists all SimpleScalar instructions with their
opcode, assembler format, and semantics. The semantics are
expressed as a C-style expression that uses the extended operators and operands described in Table 3. Operands that are not
listed in Table 3 refer to actual instruction fields described in
Figure 3. For each instruction, the next PC value (NPC) defaults
to the current PC value plus eight (CPC+8) unless otherwise
specified.

BEQ:
Opcode:
Format:
Semantics:

Branch if equal.
0x05
BEQ rs,rt,offset

BNE:
Opcode:
Format:
Semantics:

Branch if not equal.

0x06
BEQ rs,rt,offset

BLEZ:
Opcode:
Format:
Semantics:

Branch if less than or equal to zero.

0x07
BLEZ rs,offset

BGTZ:
Opcode:
Format:
Semantics:

Branch if greater than zero.

0x08
BGTZ rs,offset

BLTZ:
Opcode:
Format:
Semantics:

Branch if less than zero.

0x09
BLTZ rs,offset

BGEZ:
Opcode:
Format:
Semantics:

Branch if greater than or equal to zero.

0x0a
BGEZ rs,offset

BC1F:

Branch on floating point compare false.

if (GPR(RS) == GPR(RT))
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

if (GPR(RS) != GPR(RT))
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

if (GPR(RS) <= 0)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

if (GPR(RS) > 0)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

if (GPR(RS) < 0)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

A.1 Control instructions

J:
Opcode:
Format:
Semantics:
JAL:
Opcode:
Format:

Jump to absolute address.

0x01
J target
SET_NPC((CPC & 0xf0000000) | (TARGET<<2)))

Jump to absolute address and link.

0x02
JAL target

13

if (GPR(RS) >= 0)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

Operator/operand
FS
FT
FD
UIMM
IMM
OFFSET
CPC
NPC
SET_NPC(V)
GPR(N)
SET_GPR(N,V)
FPR_F(N)
SET_FPR_F(N,V)
FPR_D(N)
SET_FPR_D(N,V)
FPR_L(N)
SET_FPR_L(N,V)
HI
SET_HI(V)
LO
SET_LO(V)
READ_SIGNED_BYTE(A)
READ_UNSIGNED_BYTE(A)
WRITE_BYTE(V,A)
READ_SIGNED_HALF(A)
READ_UNSIGNED_HALF(A)
WRITE_HALF(V,A)
READ_WORD(A)
WRITE_WORD(V,A)
TALIGN(T)
FPALIGN(N)
OVER(X,Y)
UNDER(X,Y)
DIV0(V)

Semantics
same as field RS
same as field RT
same as field RD
IMM field unsigned-extended to word value
IMM field sign-extended to word value
IMM field sign-extended to word value
PC value of executing instruction
next PC value
Set next PC to value V
General purpose register N
Set general purpose register N to value V
Floating point register N single-precision value
Set floating point register N to single-precision value V
Floating point register N double-precision value
Set floating point register N to double-precision value V
Floating point register N literal word value
Set floating point register N to literal word value V
High result register value
Set high result register to value V
Low result register value
Set low result register to value V
Read signed byte from address A
Read unsigned byte from address A
Write byte value V at address A
Read signed half from address A
Read unsigned half from address A
Write half value V at address A
Read word from address A
Write word value V at address A
Check target T is aligned to 8 byte boundary
Check register N is wholly divisible by 2
Check for overflow when adding X to Y
Check for overflow when subtraction Y from X
Check for division by zero error with divisor V

Table 3: Operator/operand semantics

Opcode:
Format:
Semantics:

BC1T:
Opcode:
Format:
Semantics:

Semantics:

0x0b
BC1F offset
if (!FCC)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

LBU:
Opcode:
Format:
Semantics:

Load byte unsigned, displaced addressing.

0x22
LBU rt,offset(rs) inc_dec

LBU:
Opcode:
Format:
Semantics:

Load byte unsigned, indexed addressing.

0xc1
LBU rt,(rs+rd) inc_dec

SET_GPR(RT, READ_SIGNED_BYTE(GPR(RS)
+ OFFSET))

LH:
Opcode:
Format:
Semantics:

Load half signed, displaced addressing.

0x24
LH rt,offset(rs) inc_dec

Load byte signed, indexed addressing.

0xc0
LB rt,(rs+rd) inc_dec

LH:
Opcode:

Load half signed, indexed addressing.

0xc2

Branch on floating point compare true.

0x0c
BC1T offset
if (FCC)
SET_NPC(CPC + 8 + (OFFSET << 2))
else
SET_NPC(CPC + 8)

A.2 Load/store instructions

LB:
Opcode:
Format:
Semantics:

LB:
Opcode:
Format:

SET_GPR(RT,
READ_SIGNED_BYTE(GPR(RS)+GPR(RD)))

Load byte signed, displaced addressing.

0x20
LB rt,offset(rs) inc_dec

14

SET_GPR(RT,
READ_UNSIGNED_BYTE(GPR(RS)+OFFSET))

SET_GPR(RT,
READ_UNSIGNED_BYTE(GPR(RS)+GPR(RD)
))

SET_GPR(RT,
READ_SIGNED_HALF(GPR(RS)+OFFSET))

Format:
Semantics:

LHU:
Opcode:
Format:
Semantics:

SET_GPR(RT,
READ_SIGNED_HALF(GPR(RS)+GPR(RD)))

Load half unsigned, displaced addressing.

0x26
LHU rt,offset(rs) inc_dec

Load half unsigned, indexed addressing.

0xc3
LHU rt,(rs+rd) inc_dec

LW:
Opcode:
Format:
Semantics:

Load word, displaced addressing.

0x28
LW rt,offset(rs) inc_dec

LW:
Opcode:
Format:
Semantics:

Load word, indexed addressing.

0xc4
LW rt,(rs+rd) inc_dec

DLW:
Opcode:
Format:
Semantics:

Double load word, displaced addressing.

0x29
DLW rt,offset(rs) inc_dec

DLW:
Opcode:
Format:
Semantics:

Double load word, indexed addressing.

0xce
DLW rt,(rs+rd) inc_dec

Opcode:
Format:
Semantics:

L.S:
Opcode:
Format:
Semantics:

L.D:
Opcode:

L.D:

SET_GPR(RT,
READ_UNSIGNED_HALF(GPR(RS)+OFFSET))

LHU:
Opcode:
Format:
Semantics:

L.S:

Format:
Semantics:

LH rt,(rs+rd) inc_dec

Opcode:
Format:
Semantics:

SET_GPR(RT,
READ_UNSIGNED_HALF(GPR(RS)+GPR(RD)
))

SET_GPR(RT, READ_WORD(GPR(RS)+OFFSET))

SET_GPR(RT,
READ_WORD(GPR(RS)+GPR(RD)))

SET_GPR(RT, READ_WORD(GPR(RS)+OFFSET))
SET_GPR(RT+1,
READ_WORD(GPR(RS)+OFFSET+4))

SET_GPR(RT,
READ_WORD(GPR(RS)+GPR(RD)))
SET_GPR(RT+1,
READ_WORD(GPR(RS)+GPR(RD)+4))

Load word into floating point register file,

displaced addressing.
0x2a
L.S ft,offset(rs) inc_dec
SET_FPR_L(FT, READ_WORD(GPR(RS)+OFFSET))

Load word into floating point register file,

indexed addressing.
0xc5
L.S ft,(rs+rd) inc_dec
SET_FPR_L(RT,
READ_WORD(GPR(RS)+GPR(RD)))

Load double word into floating point register

file, displaced addressing.
0x2b

15

L.D ft,offset(rs) inc_dec

SET_FPR_L(FT, READ_WORD(GPR(RS)+OFFSET))
SET_FPR_L(FT+1,
READ_WORD(GPR(RS)+OFFSET+4))

Load double word into floating point register

file, indexed addressing.
0xcf
L.D ft,(rs+rd) inc_dec
SET_FPR_L(RT,
READ_WORD(GPR(RS)+GPR(RD)))
SET_FPR_L(RT+1,
READ_WORD(GPR(RS)+GPR(RD)+4))

LWL:
Opcode:
Format:
Semantics:

Load word left, displaced addressing.

0x2c
LWL offset(rs)
See ss.def for a detailed description of this
instructions semantics. NOTE: LWL does not
support pre-/post- inc/dec.

LWR:
Opcode:
Format:
Semantics:

Load word right, displaced addressing.

0x2d
LWR offset(rs)
See ss.def for a detailed description of this
instructions semantics. NOTE: LWR does not
support pre-/post- inc/dec.

SB:
Opcode:
Format:
Semantics:

Store byte, displaced addressing.

0x30
SB rt,offset(rs) inc_dec

SB:
Opcode:
Format:
Semantics:

Store byte, indexed addressing.

0xc6
SB rt,(rs+rd) inc_dec

SH:
Opcode:
Format:
Semantics:

Store half, displaced addressing.

0x32
SH rt,offset(rs) inc_dec

SH:
Opcode:
Format:
Semantics:

Store half, indexed addressing.

0xc7
SH rt,(rs+rd) inc_dec

SW:
Opcode:
Format:
Semantics:

Store word, displaced addressing.

0x34
SW rt,offset(rs) inc_dec

SW:
Opcode:
Format:
Semantics:

Store word, indexed addressing.

0xc8
SW rt,(rs+rd) inc_dec

DSW:
Opcode:
Format:
Semantics:

Double store word, displaced addressing.

0x35
DSW rt,offset(rs) inc_dec

WRITE_BYTE(GPR(RT), GPR(RS)+OFFSET)

WRITE_BYTE(GPR(RT), GPR(RS)+GPR(RD))

WRITE_HALF(GPR(RT), GPR(RS)+OFFSET)

WRITE_HALF(GPR(RT), GPR(RS)+GPR(RD))

WRITE_WORD(GPR(RT), GPR(RS)+OFFSET)

WRITE_WORD(GPR(RT), GPR(RS)+GPR(RD))

WRITE_WORD(GPR(RT), GPR(RS)+OFFSET)

Format:
Semantics:

WRITE_WORD(GPR(RT+1), GPR(RS)+OFFSET+4)

DSW:
Opcode:
Format:
Semantics:

Double store word, indexed addressing.

0xd0
DSW rt,(rs+rd) inc_dec

Double store zero, displaced addressing.

0x38
DSW rt,offset(rs) inc_dec

DSZ:
Opcode:
Format:
Semantics:

Double store zero, indexed addressing.

0xd1
DSW rt,(rs+rd) inc_dec

S.S:
Opcode:
Format:
Semantics:

S.S:
Opcode:
Format:
Semantics:

S.D:
Opcode:
Format:
Semantics:

S.D:
Opcode:
Format:
Semantics:

SWL:
Opcode:
Format:
Semantics:

SWR:
Opcode:

A.3 Integer instructions

WRITE_WORD(GPR(RT), GPR(RS)+GPR(RD))
WRITE_WORD(GPR(RT+1),
GPR(RS)+GPR(RD)+4)

DSZ:
Opcode:
Format:
Semantics:

SWR rt,offset(rs)
See ss.def for a detailed description of this
instructions semantics. NOTE: SWR does not
support pre-/post- inc/dec.

WRITE_WORD(0, GPR(RS)+OFFSET)
WRITE_WORD(0, GPR(RS)+OFFSET+4)

WRITE_WORD(0, GPR(RS)+GPR(RD))
WRITE_WORD(0, GPR(RS)+GPR(RD)+4)

Store word from floating point register file,

displaced addressing.
0x36
S.S ft,offset(rs) inc_dec
WRITE_WORD(FPR_L(FT), GPR(RS)+OFFSET)

Store word from floating point register file,

indexed addressing.
0xc9
S.S ft,(rs+rd) inc_dec
WRITE_WORD(FPR_L(FT),
GPR(RS)+GPR(RD))

Store double word from floating point register file, displaced addressing.
0x37
S.D ft,offset(rs) inc_dec
WRITE_WORD(FPR_L(FT), GPR(RS)+OFFSET)
WRITE_WORD(FPR_L(FT+1), GPR(RS)+OFFSET+4)

Store double word from floating point register file, indexed addressing.
0xd2
S.D ft,(rs+rd) inc_dec
WRITE_WORD(FPR_L(FT),
GPR(RS)+GPR(RD))
WRITE_WORD(FPR_L(FT+1),
GPR(RS)+GPR(RD)+4)

Store word left, displaced addressing.

0x39
SWL rt,offset(rs)
See ss.def for a detailed description of this
instructions semantics. NOTE: SWL does not
support pre-/post- inc/dec.
Store word right, displaced addressing.
0x3a

16

ADD:
Opcode:
Format:
Semantics:

Add signed (with overflow check).

0x40
ADD rd,rs,rt

ADDI:
check).
Opcode:
Format:
Semantics:

Add immediate signed (with overflow

OVER(GPR(RT),GPR(RT))
SET_GPR(RD, GPR(RS) + GPR(RT))

0x41
ADDI rd,rs,rt
OVER(GPR(RS),IMM)
SET_GPR(RT, GPR(RS) + IMM)

ADDU:
Opcode:
Format:
Semantics:

Add unsigned (no overflow check).

0x42
ADDU rd,rs,rt

ADDIU:
check).
Opcode:
Format:
Semantics:

Add immediate unsigned (no overflow

SET_GPR(RD, GPR(RS) + GPR(RT))

0x43
ADDIU rd,rs,rt
SET_GPR(RT, GPR(RS) + IMM)

SUB:
Opcode:
Format:
Semantics:

Subtract signed (with underflow check).

0x44
SUB rd,rs,rt

SUBU:
check).
Opcode:
Format:
Semantics:

Subtract unsigned (without underflow

UNDER(GPR(RS),GPR(RT))
SET_GPR(RD, GPR(RS) - GPR(RT))

0x45
SUBU rd,rs,rt
SET_GPR(RD, GPR(RS) - GPR(RT))

MULT:
Opcode:
Format:
Semantics:

Multiply signed.
0x46
MULT rs,rt

MULTU:
Opcode:
Format:
Semantics:

Multiply unsigned.
0x47
MULTU rs,rt

DIV:
Opcode:
Format:
Semantics:

Divide signed.
0x48
DIV rs,rt

SET_HI((RS * RT) / (1<<32))

SET_LO((RS * RT) % (1<<32))

SET_HI(((unsigned)RS * (unsigned)RT)/(1<<32))
SET_LO(((unsigned)RS*(unsigned)RT) %
(1<<32))

DIV0(GPR(RT))
SET_LO(GPR(RS) / GPR(RT))
SET_HI(GPR(RS) % GPR(RT))

DIVU
Opcode:
Format:
Semantics:

Semantics:

Divide unsigned.
0x49
DIVU rs,rt
DIV0(GPR(RT))
SET_LO((unsigned)GPR(RS)/
(unsigned)GPR(RT))
SET_HI((unsigned)GPR(RS)%(unsigned)GPR(R
T))

MFHI:
Opcode:
Format:
Semantics:

Move from HI register.

0x4a
MFHI rd

MTHI:
Opcode:
Format:
Semantics:

Move to HI register.
0x4b
MTHI rs

MFLO:
Opcode:
Format:
Semantics:

Move from LO register.

0x4c
MFLO rd

MTLO:
Opcode:
Format:
Semantics:

Move to LO register.
0x4d
MTLO rs

AND:
Opcode:
Format:
Semantics:

Logical AND.
0x4e
AND rd,rs,rt

ANDI:
Opcode:
Format:
Semantics:

Logical AND immediate.

0x4f
ANDI rd,rt,imm

OR:
Opcode:
Format:
Semantics:

Logical OR.
0x50
OR rd,rs,rt

ORI:
Opcode:
Format:
Semantics:

Logical OR immediate.
0x51
ORI rd,rt,imm

XOR:
Opcode:
Format:
Semantics:

Logical XOR.
0x52
XOR rd,rs,rt

XORI:
Opcode:
Format:
Semantics:

Logical XOR immediate.

0x53
ORI rd,rt,uimm

NOR:
Opcode:
Format:

Logical NOR.
0x54
NOR rd,rs,rt

SET_GPR(RD, HI)

SET_HI(GPR(RS))

SET_GPR(RD, LO)

SLL:
Opcode:
Format:
Semantics:

Shift left logical.

0x55
SLL rd,rt,shamt

SLLV:
Opcode:
Format:
Semantics:

Shift left logical variable.

0x56
SLLV rd,rt,rs

SRL:
Opcode:
Format:
Semantics:

Shift right logical.

0x57
SRL rd,rt,shamt

SRLV:
Opcode:
Format:
Semantics:

Shift right logical variable.

0x58
SRLV rd,rt,rs

SRA:
Opcode:
Format:
Semantics:

Shift right arithmetic.

0x59
SRA rd,rt,shamt

SRAV:
Opcode:
Format:
Semantics:

Shift right arithmetic variable.

0x59
SRAV rd,rt,rs

SLT:
Opcode:
Format:
Semantics:

Set register if less than.

0x5b
SLT rd,rs,rt

SLTI:
Opcode:
Format:
Semantics:

Set register if less than immediate.

0x5c
SLTI rd,rs,imm

SLTU:
Opcode:
Format:
Semantics:

Set register if less than unsigned.

0x5d
SLTU rd,rs,rt

SLTIU:
Opcode:
Format:
Semantics:

Set register if less than unsigned immediate.

0x5d
SLTIU rd,rs,imm

SET_LO(GPR(RS))

SET_GPR(RD, GPR(RS) & GPR(RT))

SET_GPR(RT, GPR(RS) & UIMM)

SET_GPR(RD, GPR(RS) | GPR(RT))

SET_GPR(RT, GPR(RS) | UIMM)

SET_GPR(RD, GPR(RS) ^ GPR(RT))

SET_GPR(RD, ~(GPR(RS) | GPR(RT)))

SET_GPR(RD, GPR(RT) << SHAMT)

SET_GPR(RD, GPR(RT) << (GPR(RS) & 0x1f))

SET_GPR(RD, GPR(RT) >> SHAMT)

SET_GPR(RD, GPR(RT) << (GPR(RS) & 0x1f))

SET_GPR(RD, SEX(GPR(RT) >> SHAMT, 31 SHAMT))

SET_GPR(RD, SEX(GPR(RT) >> SHAMT, 31 (GPR(RD) & 0x1f)))

SET_GPR(RD, (GPR(RS) < GPR(RT)) ? 1 : 0)

SET_GPR(RD, (GPR(RS) < IMM) ? 1 : 0)

SET_GPR(RD,
((unsigned)GPR(RS)<(unsigned)GPR(RT)) ? 1 : 0)

SET_GPR(RD,
((unsigned)GPR(RS)<(unsigned)GPR(RT)) ? 1 : 0)

A.4 Floating-point instructions

SET_GPR(RT, GPR(RS) ^ UIMM)

ADD.S:
Opcode:
Format:
Semantics:

17

Add floating point, single precision.

0x70
ADD.S fd,fs,ft
FPALIGN(FD)

Semantics:

FPALIGN(FS)
FPALIGN(FT)
SET_FPR_F(FD, FPR_F(FS) + FPR_F(FT)))

ADD.D:
Opcode:
Format:
Semantics:

Add floating point, double-precision.

0x71
ADD.D fd,fs,ft

SUB.S:
Opcode:
Format:
Semantics:

Subtract floating point, single precision.

0x72
SUB.S fd,fs,ft

SUB.D:
Opcode:
Format:
Semantics:

MUL.S:
Opcode:
Format:
Semantics:

MUL.D:
Opcode:
Format:
Semantics:

DIV.S:
Opcode:
Format:
Semantics:

DIV.D:
Opcode:
Format:
Semantics:

ABS.S:
Opcode:
Format:

FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
SET_FPR_D(FD, FPR_D(FS) + FPR_D(FT)))

FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
SET_FPR_F(FD, FPR_F(FS) - FPR_F(FT)))

Subtract floating point, double precision.

0x73
SUB.D fd,fs,ft
FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
SET_FPR_D(FD, FPR_D(FS) - FPR_D(FT)))

Multiply floating point, single precision.

0x74
MUL.S fd,fs,ft
FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
SET_FPR_F(FD,FPR_F(FS)*FPR_F(FT)))

Multiply floating point, double precision.

0x75
MUL.D fd,fs,ft
FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
SET_FPR_D(FD, FPR_D(FS) * FPR_D(FT)))

Divide floating point, single precision.

0x76
DIV.S fd,fs,ft
FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
DIV0(FPR_F(FT))
SET_FPR_F(FD, FPR_F(FS) / FPR_F(FT)))

Divide floating point, double precision.

0x77
DIV.D fd,fs,ft
FPALIGN(FD)
FPALIGN(FS)
FPALIGN(FT)
DIV0(FPR_D(FT))
SET_FPR_D(FD, FPR_D(FS) / FPR_D(FT)))

Absolute value, single precision.

0x78
ABS.S fd,fs

18

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_F(FD, fabs((double)FPR_F(FS))))

ABS.D:
Opcode:
Format:
Semantics:

Absolute value, double precision.

0x79
ABS.D fd,fs

MOV.S:
Opcode:
Format:
Semantics:

Move floating point value, single precision.

0x7a
MOV.S fd,fs

MOV.D:
Opcode:
Format:
Semantics:

Move floating point value, double precision.

0x7b
MOV.D fd,fs

NEG.S:
Opcode:
Format:
Semantics:

Negate floating point value, single precision.

0x7c
NEG.S fd,fs

NEG.D:
sion.
Opcode:
Format:
Semantics:

Negate floating point value, double preci-

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD, fabs(FPR_D(FS))))

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_F(FD, FPR_F(FS))

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD, FPR_D(FS))

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_F(FD, -FPR_F(FS))

0x7d
NEG.D fd,fs
FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD, -FPR_D(FS))

CVT.S.D:
Opcode:
Format:
Semantics:

Convert double precision to single precision.

0x80
CVT.S.D fd,fs

CVT.S.W:
Opcode:
Format:
Semantics:

Convert integer to single precision.

0x81
CVT.S.W fd,fs

CVT.D.S:
Opcode:
Format:
Semantics:

Convert single precision to double precision.

0x82
CVT.D.S fd,fs

CVT.D.W:
Opcode:
Format:

Convert integer to double precision.

0x83
CVT.D.W fd,fs

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD, -FPR_D(FS))

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_F(FD, (float)FPR_L(FS))

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD,(double)FPR_F(FS))

Semantics:

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD,(double)FPR_L(FS))

CVT.W.S:
Opcode:
Format:
Semantics:

Convert single precision to integer.

0x84
CVT.W.S fd,fs

CVT.W.D:
Opcode:
Format:
Semantics:

Convert double precision to integer.

0x85
CVT.W.D fd,fs

C.EQ.S:
Opcode:
Format:
Semantics:

Test if equal, single precision.

0x90
C.EQ.S fs,ft

FPALIGN(FS)
SET_FPR_F(FD,sqrt((double)FPR_F(FS)))

SQRT.D:
Opcode:
Format:
Semantics:

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_L(FD, (long)FPR_F(FS))

Square root, double precision.

0x97
SQRT.D fd,fs
FPALIGN(FD)
FPALIGN(FS)
SET_FPR_D(FD, sqrt(FPR_D(FS)))

A.5 Miscellaneous instructions

C.EQ.D:
Opcode:
Format:
Semantics:

C.LT.S:
Opcode:
Format:
Semantics:

FPALIGN(FD)
FPALIGN(FS)
SET_FPR_L(FD, (long)FPR_D(FS))

FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_F(FS) == FPR_F(FT))

Test if equal, double precision.

0x91
C.EQ.D fs,ft
FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_D(FS) == FPR_D(FT))

Test if less than, single precision.

0x92
C.LT.S fs,ft
FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_F(FS) < FPR_F(FT))

C.LT.D:
Opcode:
Format:
Semantics:

Test if less than, double precision.

0x93
C.LT.D fs,ft

C.LE.S:
Opcode:
Format:
Semantics:

Test if less than or equal, single precision.

0x94
C.LE.S fs,ft

C.LE.D:
Opcode:
Format:
Semantics:

Test if less than or equal, double precision.

0x95
C.LE.D fs,ft

SQRT.S:
Opcode:
Format:
Semantics:

NOP:
Opcode:
Format:
Semantics:

No operation.
0x00
NOP
None

SYSCALL:
Opcode:
Format:
Semantics:

System call.
0xa0
SYSCALL
See Appendix B for details

BREAK:
Opcode:
Format:
Semantics:

Declare a program error.

0xa1
BREAK uimm
Actions are simulator-dependent. Typically,
an error message is printed and abort() is
called.

LUI:
Opcode:
Format:
Semantics:

Load upper immediate.

0xa2
LUI uimm

MFC1:

Move from floating point to integer register

file.
0xa3
MFC1 rt,fs

Opcode:
Format:
Semantics:
MTC1:

FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_D(FS) < FPR_D(FT))

Opcode:
Format:
Semantics:

SET_GPR(RT, UIMM << 16)

SET_GPR(RT, FPR_L(FS))

Move from integer to floating point register

file.
0xa5
MTC1 rt,fs
SET_FPR_L(FS, GPR(RT))

B System call definitions

This appendix lists all system calls supported by the simulators with their system call code (syscode), interface specification,
and appropriate POSIX Unix reference. Systems calls are initiated with the SYSCALL instruction. Prior to execution of a
SYSCALL instruction, register $v0 should be loaded with the
system call code. The arguments of the system call interface prototype should be loaded into registers $a0 - $a3 in the order specified by the system call interface prototype, e.g., for:

FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_F(FS) <= FPR_F(FT))

FPALIGN(FS)
FPALIGN(FT)
SET_FCC(FPR_D(FS) <= FPR_D(FT))

read(int fd, char *buf, int nbyte),

0x03 is loaded into $v0, fd is loaded into $a0, buf into $a1, and
nbyte into $a2.

Square root, single precision.

0x96
SQRT.S fd,fs

EXIT:

FPALIGN(FD)

19

Exit process.

Syscode:
Interface:
Semantics:

0x01
void exit(int status);
See exit(2).

READ:
Syscode:
Interface:
Semantics:

Read from file to buffer.

0x03
int read(int fd, char *buf, int nbyte);
See read(2).

WRITE:
Syscode:
Interface:
Semantics:

Write from a buffer to a file.

0x04
int write(int fd, char *buf, int nbyte);
See write(2).

OPEN:
Syscode:
Interface:
Semantics:

Open a file.
0x05
int open(char *fname, int flags, int mode);
See open(2).

CLOSE:
Syscode:
Interface:
Semantics:

Close a file.
0x06
int close(int fd);
See close(2).

CREAT:
Syscode:
Interface:
Semantics:

Create a file.
0x08
int creat(char *fname, int mode);
See creat(2).

UNLINK:
Syscode:
Interface:
Semantics:

Delete a file.
0x0a
int unlink(char *fname);
See unlink(2).

CHDIR:
Syscode:
Interface:
Semantics:

Change process directory.

0x0c
int chdir(char *path);
See chdir(2).

CHMOD:
Syscode:
Interface:
Semantics:

Change file permissions.

0x0f
int chmod(int *fname, int mode);
See chmod(2).

CHOWN:
Syscode:
Interface:
Semantics:

Change file owner and group.

0x10
int chown(char *fname, int owner, int group);
See chown(2).

BRK:
Syscode:
Interface:
Semantics:

Change process break address.

0x11
int brk(long addr);
See brk(2).

LSEEK:
Syscode:
Interface:
Semantics:

Move file pointer.

0x13
long lseek(int fd, long offset, int whence);
See lseek(2).

GETPID:
Syscode:
Interface:

Get process identifier.

0x14
int getpid(void);

Semantics:

See getpid(2).

GETUID:
Syscode:
Interface:
Semantics:

Get user identifier.

0x18
int getuid(void);
See getuid(2).

ACCESS:
Syscode:
Interface:
Semantics:

Determine accessibility of a file.

0x21
int access(char *fname, int mode);
See access(2).

STAT:
Syscode:
Interface:

Get file status.

0x26
struct stat
{
short
st_dev;
long
st_ino;
unsigned short st_mode;
short
st_nlink;
short
st_uid;
short
st_gid;
short
st_rdev;
int
st_size;
int
st_atime;
int
st_spare1;
int
st_mtime;
int
st_spare2;
int
st_ctime;
int
st_spare3;
long
st_blksize;
long
st_blocks;
long
st_gennum;
long
st_spare4;
};
int stat(char *fname, struct stat *buf);
See stat(2).

Semantics:

20

LSTAT:
Syscode:
Interface:
Semantics:

Get file status (and dont dereference links).

0x28
int lstat(char *fname, struct stat *buf);
See lstat(2).

DUP:
Syscode:
Interface:
Semantics:

Duplicate a file descriptor.

0x29
int dup(int fd);
See dup(2).

PIPE:
Syscode:
Interface:
Semantics:

Create an interprocess comm. channel.

0x2a
int pipe(int fd[2]);
See pipe(2).

GETGID:
Syscode:
Interface:
Semantics:

Get group identifier.

0x2f
int getgid(void);
See getgid(2).

IOCTL:
Syscode:
Interface:
Semantics:

Device control interface.

0x36
int ioctl(int fd, int request, char *arg);
See ioctl(2).

FSTAT:
Syscode:
Interface:
Semantics:

Get file descriptor status.

0x3e
int fstat(int fd, struct stat *buf);
See fstat(2).

GETPAGESIZE:
Syscode:
Interface:
Semantics:

Get page size.

0x40
int getpagesize(void);
See getpagesize(2).

Syscode:
Interface:
Semantics:

GETDTABLESIZE: Get file descriptor table size.

Syscode:
0x59
Interface:
int getdtablesize(void);
Semantics:
See getdtablesize(2).
DUP2:
Syscode:
Interface:
Semantics:

Duplicate a file descriptor.

0x5a
int dup2(int fd1, int fd2);
See dup2(2).

FCNTL:
Syscode:
Interface:
Semantics:

File control.
0x5c
int fcntl(int fd, int cmd, int arg);
See fcntl(2).

SELECT:
Syscode:
Interface:

Synchronous I/O multiplexing.

0x5d
int select (int width, fd_set *readfds, fd_set
*writefds, fd_set *exceptfds, struct timeval
*timeout);
See select(2).

Semantics:

GETTIMEOFDAY: Get the date and time.

Syscode:
0x74
Interface:
struct timeval {
long tv_sec;
long tv_usec;
};
struct int {
timezone tz_minuteswest;
int tz_dsttime;
};
int gettimeofday(struct timeval *tp,
struct timezone *tzp);
Semantics:
See gettimeofday(2).
WRITEV:
Syscode:
Interface:
Semantics:

Write output, vectored.

0x79
int writev(int fd, struct iovec *iov, int cnt);
See writev(2).

UTIMES:
Syscode:
Interface:
Semantics:

Set file times.

0x8a
int utimes(char *file, struct timeval *tvp);
See utimes(2).

GETRLIMIT:
Syscode:
Interface:
Semantics:

Get maximum resource consumption.

0x90
int getrlimit(int res, struct rlimit *rlp);
See getrlimit(2).

SETRLIMIT:

Set maximum resource consumption.

21

0x91
int setrlimit(int res, struct rlimit *rlp);
See setrlimit(2).